Jeffy packages pretrained classifiers for jobs where you already know the labels: banking support intent, SMS spam, sentiment, news topics, and multilingual voice-command intent. Rather than asking a general-purpose LLM to decide each time, it turns the input into an embedding and applies a small logistic-regression classifier. The source repository was created October 2, and its first PyPI alpha was uploaded that day; this is a new project, not a newly discovered old one.

  • Try it: pip install jeffy-classify then jeffy-serve for a local interface at localhost:8400. POST {"text":"I was charged twice","task":"banking77"} to /v1/predict, or use Engine.predict() in Python. Train a classifier on labeled CSV, TSV, or JSONL with jeffy-train.
  • What ships: The current README lists 68 task classifiers, including 51 language-specific intent classifiers. The Show HN post says 13, so treat the repository’s current catalog—not the launch blurb—as the moving target.
  • Cost and footprint: The code is MIT-licensed and runs locally without a GPU or hosted API fee. The English encoder downloads about 1.2 GB on first use; the README estimates roughly 2 GB runtime memory and 50–80 ms per prediction on its tested ARM Linux CPU. The classifier head is small, but the whole stack is not a tiny dependency.
  • Limits: Tasks need an existing trained head; there is no automatic task choice, zero-shot general classifier or LLM fallback. Reported benchmark scores come from the project’s own evaluations, not independent testing. Its author also says redistribution permissions for all dataset-derived classifier weights have not been independently verified; review the attribution file before commercial redistribution.

Useful if you have repeated, well-defined routing decisions and want a local alternative to per-item LLM calls. Not a substitute for open-ended classification or a benchmark of your own data.